Deep reinforcement learning for traffic signal control often optimizes aggregate efficiency while leaving a subset of vehicles with excessive waiting. This paper proposes a fairness-aware PPO controller that adds lane-level long-wait vehicle counts to the state and introduces an excess-wait penalty into the reward. A coefficient γ controls the efficiency–fairness trade-off. In SUMO experiments on an isolated intersection, PPO-Ours (γ = 0.20) reduces maximum waiting time from 688.33 s to 480.23 s and waiting-time standard deviation from 158.87 s to 98.45 s relative to PPO-Basic, while preserving competitive average-queue performance. Ablation results show that γ provides a practical trade-off knob rather than a universally optimal setting under all metrics.